EEG Classification¶

Scratch¶

Group 11: Yazid Mouline & Guillaume Requena | AML Class 2020¶

from google.colab import drive
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns
from time import time

import codecs
from IPython.display import Image
import networkx as nx

from scipy import stats

import sys
sys.path.append('/content/drive/My Drive/AML/Chal3 - EEG Classification/Code/')


import shutil
import os
from sklearn.model_selection import train_test_split

import shutil
from sklearn.model_selection import StratifiedKFold

!pip install -U mne

Requirement already up-to-date: mne in /usr/local/lib/python3.6/dist-packages (0.20.5)
Requirement already satisfied, skipping upgrade: numpy>=1.11.3 in /usr/local/lib/python3.6/dist-packages (from mne) (1.18.4)
Requirement already satisfied, skipping upgrade: scipy>=0.17.1 in /usr/local/lib/python3.6/dist-packages (from mne) (1.4.1)

Data importation¶

drive.mount('/content/drive/',force_remount=True)

Mounted at /content/drive/

params = {'resample_to': 1000,
                 'D': 3,
                 'F1': 12,
                 'dropoutRate1': 0.52,
                 'dropoutRate2': 0.36,
                 'lr': 0.00066,
                 'norm_rate': 0.275
                 }

import sys
sys.path.append('/content/drive/My Drive/AML/Chal3 - EEG Classification/Code')
from data import DataBuildClassifier

data_loader = DataBuildClassifier('/content/drive/My Drive/AML/Chal3 - EEG Classification/Data/eeg') #Path to directory with data (i.e NewData contatins 25/, 26/ ....)
all_subjects = [25, 26,27,28,29,30, 32,33,34,35,36,37,38]
subjects = data_loader.get_data(all_subjects,shuffle=False, windows=[(0.2,0.5)],baseline_window=(0.2,0.3),resample_to=params['resample_to'])

Data exploration¶

1.1 Study on one individual¶

print(subjects.keys())
X,y = subjects[25]
print(X.shape) #EEG epochs (Trials) x Time x Channels
print(y.shape)

dict_keys([25, 26, 27, 28, 29, 30, 32, 33, 34, 35, 36, 37, 38])
(329, 300, 19)
(329,)

X[0,:,16].shape

(300,)

def plot_mean_amp(X, y, epoch, subplot):
  plot = []
  for i in range(len(X[0])):
    plot.append(np.mean(X[epoch, i, :]))

  maxi = np.round(np.max(plot), 2)
  argmaxi = np.argmax(plot)
  mini = np.round(np.min(plot), 2)
  argmini = np.argmin(plot)

  if (y[epoch]==1):
    label = 'Non-Target'
  else :
    label = 'Target'
  
  plt.subplot(subplot)

  plt.axhline(maxi, color='brown', linestyle='--')
  plt.axhline(mini, color='orange', linestyle='--')
  plt.axvline(argmaxi, color='brown', linestyle='-')
  plt.axvline(argmini, color='orange', linestyle='-')

  plt.xlabel('Time (ms)')
  plt.ylabel('Amplitude (V)')
  plt.title('Mean amplitude over the 19 channels for the epoch {} of class {}'.format(epoch, label))

  plt.legend({'Max = {}V'.format(maxi):maxi,
              'Argmax = {}ms'.format(argmaxi):argmaxi,
              'Min = {}V'.format(mini):mini, 
              'Argmin = {}ms'.format(argmini):argmini},
            loc='upper center', bbox_to_anchor=(0.5, 0.9));

  plt.plot(plot)

fig = plt.figure(figsize=(25,14))

plot_mean_amp(X, y, epoch=50, subplot=221)
plot_mean_amp(X, y, epoch=150, subplot=222)
plot_mean_amp(X, y, epoch=250, subplot=223)
plot_mean_amp(X, y, epoch=300, subplot=224)

plt.show()

As we can see on the chart right above, there is always a peak at the beginning of class Target and then it decreases. This is probably where we can interpret the signal as Target label. However it seems to be more random when the signal is labeled as Non-Target so when the person is having a look a the game without making any decisions.

def plot_mean_peak(X):
  argmaxis_NT = []
  argmaxis_T = []

  for epoch in range(len(X)):
    plot = []
    for i in range(len(X[0])):
      plot.append(np.mean(X[epoch, i, :]))

    if (y[epoch]==1): #Non-Target
      argmaxis_NT.append(np.argmax(plot))
    else :            #Target
      argmaxis_T.append(np.argmax(plot))

  return argmaxis_NT, argmaxis_T

def plot_peak_distrib(argmaxis, channel='None', one_channel=False, target=True):
  f, (ax_box, ax_hist) = plt.subplots(2, sharex=True, gridspec_kw= {"height_ratios": (0.2, 1)}, figsize=(15,7))
  if (target):
    if (one_channel):
      plt.title('Distribution of peak index for Target EEG at channel {}'.format(channel))
    else : 
      plt.title('Distribution of peak index for Target EEG')
  else : 
    if (one_channel):
      plt.title('Distribution of peak index for Non-Target EEG at channel {}'.format(channel))
    else : 
      plt.title('Distribution of peak index for Non-Target EEG')

  argmaxis = pd.Series(argmaxis)

  mean = round(argmaxis.mean(), 2)
  median = round(argmaxis.median(), 2)
  mode = round(argmaxis.mode()[0], 2)
  maxi = round(argmaxis.max(), 2)
  mini = round(argmaxis.min(), 2)

  sns.boxplot(argmaxis, ax=ax_box)
  ax_box.axvline(mean, color='r', linestyle='--')
  ax_box.axvline(median, color='g', linestyle='-')
  ax_box.axvline(mode, color='b', linestyle='-')
  ax_box.axvline(maxi, color='brown', linestyle='-')
  ax_box.axvline(mini, color='orange', linestyle='-')

  sns.distplot(argmaxis, ax=ax_hist, bins=20)
  ax_hist.axvline(mean, color='r', linestyle='--')
  ax_hist.axvline(median, color='g', linestyle='-')
  ax_hist.axvline(mode, color='b', linestyle='-')
  ax_hist.axvline(maxi, color='brown', linestyle='-')
  ax_hist.axvline(mini, color='orange', linestyle='-')
  
  plt.xlabel('Time (ms)')
  plt.legend({'Mean={}'.format(mean):mean,
              'Median={}'.format(median):median,
              'Mode={}'.format(mode):mode, 
              'Max={}'.format(maxi):maxi,
              'Min={}'.format(mini):mini},
            loc='upper right')

  ax_box.set(xlabel='')

argmaxis_NT, argmaxis_T = plot_mean_peak(X)

plot_peak_distrib(argmaxis_T)
plot_peak_distrib(argmaxis_NT, target=False)

The remarks we had previously appears to be quite wrong according to the fact in average over all the channels there is no peak at the beginning of a Target sequence. Actually it could be linked to the fact that in average the brain is implicated every time. While in the Non-Target sequence we can observe a peak a the beginning in average which could be related to the first glance on the game and all the understanding you need to do when you first face a game. Then the amplitude turns to decrease because you understand what you have to do so you don't have to think that much.

Now we will try to understand where are the most solicitated part of the brain according to the EEG results.

def plot_EEG_cmap(info, X, y, epoch, time, subplot):
  pos_list = dict()

  for i in range(1,len(info)+1):
    theta = np.radians(info.loc[i,'theta']+90)
    x = np.cos(theta)*info.loc[i,'radius']
    Y = np.sin(theta)*info.loc[i,'radius']
    pos_list[i]=(x,Y)

  node_labels = {i:info.loc[i,'class'].replace(' ','') for i in range(1,len(info)+1)}

  if (y[epoch]==1):
    label = 'Non-Target'
  else :
    label = 'Target'
  
  plt.subplot(subplot)
  #plt.text(x=0, y=0.33, s='Nasion', horizontalalignment='center');
  plt.text(x=0, y=-0.58, s='Inion', horizontalalignment='center', fontsize=13);


  g = nx.Graph()
  g.add_nodes_from(pos_list.keys())
  for i in pos_list.keys():
      g.nodes[i]['amp'] = X[epoch, time, i-1]

  colors = list(nx.get_node_attributes(g,'amp').values())

  nc = nx.draw_networkx_nodes(g, pos_list, labels = node_labels, font_color= 'w', node_color=colors, node_size = 600, cmap=plt.cm.jet)
  nx.draw_networkx_labels(g, pos_list, labels = node_labels, font_color= 'w', font_size=10, font_weight= 'bold')

  cbar = plt.colorbar(nc)
  cbar.set_label('Amplitude (V)', rotation=270, fontsize=10, labelpad=15)
  plt.axis('off')
  plt.title('EEG of the epoch {} at {}ms labeled as {} \n\n Nasion'.format(epoch, time, label))

info = pd.read_csv('/content/drive/My Drive/AML/Chal3 - EEG Classification/Data/eeg/order_locations.info',
                   sep ='\t',
                   header = 0,
                   names = ['theta', 'radius', 'class'])

fig = plt.figure(figsize=(30,7))

#EEG at epoch 0 for time 100ms, 150ms and 250ms
plot_EEG_cmap(info, X, y, 0, 100, 131)
plot_EEG_cmap(info, X, y, 0, 150, 132)
plot_EEG_cmap(info, X, y, 0, 250, 133)

plt.show()

fig = plt.figure(figsize=(30,7))

#EEG at epoch 150 for time 100ms, 150ms and 250ms
plot_EEG_cmap(info, X, y, 150, 100, 131)
plot_EEG_cmap(info, X, y, 150, 150, 132)
plot_EEG_cmap(info, X, y, 150, 250, 133)

plt.show()

For the Non-Target sequence we can observe that F4 and F3 are the most solicitated in the first place. Then we can see that PO8 and PO7 are the most active.

fig = plt.figure(figsize=(30,7))

#EEG at epoch 300 for time 100ms, 150ms and 250ms
plot_EEG_cmap(info, X, y, 300, 100, 131)
plot_EEG_cmap(info, X, y, 300, 150, 132)
plot_EEG_cmap(info, X, y, 300, 250, 133)

plt.show()

As we could have expected, we see a slow negative wave with a maximum in the parietooccipital region was present in each participant's averaged FRPs in the control-on conditions and was absent or had much lower amplitude in the control-off condition. So we decided to plot the distribution of the most active part of the brain.

def plot_onechannel_peak(X, node_label):  #node_label must be a string
  map_node_labels = {1: 'Fz',
               2: 'F3',
               3: 'F4',
               4: 'Cz',
               5: 'C3',
               6: 'C4',
               7: 'Pz',
               8: 'P1',
               9: 'P2',
               10: 'P3',
               11: 'P4',
               12: 'POz',
               13: 'PO3',
               14: 'PO4',
               15: 'PO7',
               16: 'PO8',
               17: 'Oz',
               18: 'O1',
               19: 'O2'}
  
  inv_map_node_labels = {v: k for k, v in map_node_labels.items()}
  argmaxis_NT = []
  argmaxis_T = []
  channel_nb = inv_map_node_labels[node_label]

  for epoch in range(len(X)):
    plot = []
    for i in range(len(X[0])):
      plot.append(X[epoch, i, channel_nb])

    if (y[epoch]==1): #Non-Target
      argmaxis_NT.append(np.argmax(plot))
    else :            #Target
      argmaxis_T.append(np.argmax(plot))

  return argmaxis_NT, argmaxis_T

argmaxis_NT_F4, argmaxis_T_F4 = plot_onechannel_peak(X, 'F4')
plot_peak_distrib(argmaxis_T_F4, channel='F4', one_channel=True)
plot_peak_distrib(argmaxis_NT_F4, channel='F4', one_channel=True, target=False)

argmaxis_NT_F3, argmaxis_T_F3 = plot_onechannel_peak(X, 'F3')
plot_peak_distrib(argmaxis_T_F3, channel='F3', one_channel=True)
plot_peak_distrib(argmaxis_NT_F3, channel='F3', one_channel=True, target=False)

argmaxis_NT_PO8, argmaxis_T_PO8 = plot_onechannel_peak(X, 'PO8')
plot_peak_distrib(argmaxis_T_PO8, channel='PO8', one_channel=True)
plot_peak_distrib(argmaxis_NT_PO8, channel='PO8', one_channel=True, target=False)

argmaxis_NT_PO7, argmaxis_T_PO7 = plot_onechannel_peak(X, 'PO7')
plot_peak_distrib(argmaxis_T_PO7, channel='PO7', one_channel=True)
plot_peak_distrib(argmaxis_NT_PO7, channel='PO7', one_channel=True, target=False)

def plot_onechannel_amp(X, y, epoch, subplot, node_label):
  map_node_labels = {1: 'Fz',
               2: 'F3',
               3: 'F4',
               4: 'Cz',
               5: 'C3',
               6: 'C4',
               7: 'Pz',
               8: 'P1',
               9: 'P2',
               10: 'P3',
               11: 'P4',
               12: 'POz',
               13: 'PO3',
               14: 'PO4',
               15: 'PO7',
               16: 'PO8',
               17: 'Oz',
               18: 'O1',
               19: 'O2'}
  
  inv_map_node_labels = {v: k for k, v in map_node_labels.items()}
  channel_nb = inv_map_node_labels[node_label]

  plot = []
  for i in range(len(X[0])):
    plot.append(X[epoch, i, channel_nb])

  maxi = np.round(np.max(plot), 2)
  argmaxi = np.argmax(plot)
  mini = np.round(np.min(plot), 2)
  argmini = np.argmin(plot)

  if (y[epoch]==1):
    label = 'Non-Target'
  else :
    label = 'Target'
  
  plt.subplot(subplot)

  plt.axhline(maxi, color='brown', linestyle='--')
  plt.axhline(mini, color='orange', linestyle='--')
  plt.axvline(argmaxi, color='brown', linestyle='-')
  plt.axvline(argmini, color='orange', linestyle='-')

  plt.xlabel('Time (ms)')
  plt.ylabel('Amplitude (V)')
  plt.title('Amplitude of the {} channel for the epoch {} of class {}'.format(node_label, epoch, label))

  plt.legend({'Max = {}V'.format(maxi):maxi,
              'Argmax = {}ms'.format(argmaxi):argmaxi,
              'Min = {}V'.format(mini):mini, 
              'Argmin = {}ms'.format(argmini):argmini},
            loc='upper center', bbox_to_anchor=(0.5, 0.9));

  plt.plot(plot)

fig = plt.figure(figsize=(25,14))

node_label = 'PO7'

plot_onechannel_amp(X, y, epoch=50, subplot=221, node_label=node_label)
plot_onechannel_amp(X, y, epoch=150, subplot=222, node_label=node_label)
plot_onechannel_amp(X, y, epoch=250, subplot=223, node_label=node_label)
plot_onechannel_amp(X, y, epoch=300, subplot=224, node_label=node_label)

plt.show()

X_30,y_30 = subjects[30]
print(X_30.shape) #EEG epochs (Trials) x Time x Channels
print(y_30.shape)

fig = plt.figure(figsize=(25,14))

node_label = 'PO7'

plot_onechannel_amp(X_30, y_30, epoch=50, subplot=221, node_label=node_label)
plot_onechannel_amp(X_30, y_30, epoch=150, subplot=222, node_label=node_label)
plot_onechannel_amp(X_30, y_30, epoch=300, subplot=223, node_label=node_label)
plot_onechannel_amp(X_30, y_30, epoch=400, subplot=224, node_label=node_label)

plt.show()

(451, 300, 19)
(451,)

1.2 Study over all individuals¶

1.2.1 Temporal Information¶

print(subjects.keys())
X,y = subjects[25]

dict_keys([25, 26, 27, 28, 29, 30, 32, 33, 34, 35, 36, 37, 38])

def separate_T_NT(X,y):
  X_T = []
  X_NT = []

  y_T = []
  y_NT = []

  for epoch in range(len(X)):
    if (y[epoch]==1):
      X_NT.append(X[epoch])
      y_NT.append(y[epoch])
    else : 
      X_T.append(X[epoch])
      y_T.append(y[epoch])
  return np.array(X_T), np.array(y_T), np.array(X_NT), np.array(y_NT)

def plot_mean_amp_indiv(X, subplot, node_label, target=True):
  map_node_labels = {1: 'Fz',
               2: 'F3',
               3: 'F4',
               4: 'Cz',
               5: 'C3',
               6: 'C4',
               7: 'Pz',
               8: 'P1',
               9: 'P2',
               10: 'P3',
               11: 'P4',
               12: 'POz',
               13: 'PO3',
               14: 'PO4',
               15: 'PO7',
               16: 'PO8',
               17: 'Oz',
               18: 'O1',
               19: 'O2'}
  
  inv_map_node_labels = {v: k for k, v in map_node_labels.items()}
  channel_nb = inv_map_node_labels[node_label]
  
  plot = []
  for i in range(len(X[0])):
    plot.append(np.mean(X[:, i, channel_nb]))

  maxi = np.round(np.max(plot), 2)
  argmaxi = np.argmax(plot)
  mini = np.round(np.min(plot), 2)
  argmini = np.argmin(plot)

  if (target):
    label = 'Target'
  else :
    label = 'Non-Target'
  
  plt.subplot(subplot)

  plt.axhline(maxi, color='brown', linestyle='--')
  plt.axhline(mini, color='orange', linestyle='--')
  plt.axvline(argmaxi, color='brown', linestyle='-')
  plt.axvline(argmini, color='orange', linestyle='-')

  plt.xlabel('Time (ms)')
  plt.ylabel('Amplitude (V)')
  plt.title('Mean amplitude over all the epochs of the {} of class {} for individual 25'.format(node_label, label))

  plt.legend({'Max = {}V'.format(maxi):maxi,
              'Argmax = {}ms'.format(argmaxi):argmaxi,
              'Min = {}V'.format(mini):mini, 
              'Argmin = {}ms'.format(argmini):argmini},
            loc='upper center', bbox_to_anchor=(0.5, 0.9));

  plt.plot(plot)

X_T, y_T, X_NT, y_NT = separate_T_NT(X,y)

fig = plt.figure(figsize=(25,7))

plot_mean_amp_indiv(X_T, subplot=121, node_label='PO8')
plot_mean_amp_indiv(X_NT, subplot=122, node_label='PO8', target=False)

plt.show()

def plot_mean_amp_per_indiv(X, node_label, individual, target=True):
  map_node_labels = {1: 'Fz',
               2: 'F3',
               3: 'F4',
               4: 'Cz',
               5: 'C3',
               6: 'C4',
               7: 'Pz',
               8: 'P1',
               9: 'P2',
               10: 'P3',
               11: 'P4',
               12: 'POz',
               13: 'PO3',
               14: 'PO4',
               15: 'PO7',
               16: 'PO8',
               17: 'Oz',
               18: 'O1',
               19: 'O2'}
  
  inv_map_node_labels = {v: k for k, v in map_node_labels.items()}
  channel_nb = inv_map_node_labels[node_label]

  plot = []
  for i in range(len(X[0])):
    plot.append(np.mean(X[:, i, channel_nb]))

  maxi = np.round(np.max(plot), 2)
  argmaxi = np.argmax(plot)
  mini = np.round(np.min(plot), 2)
  argmini = np.argmin(plot)

  if (target):
    label = 'Target'
  else :
    label = 'Non-Target'
  
  plt.xlabel('Time (ms)')
  plt.ylabel('Amplitude (V)')
  plt.title('Mean amplitude on {} over all the individuals of class {}'.format(node_label, label))

  plt.plot(plot)

individuals = list(subjects.keys())
all_X_T = dict()
all_X_NT = dict()

for individual in individuals:
  X, y = subjects[individual]

  X_T, y_T, X_NT, y_NT = separate_T_NT(X, y)
  all_X_T[individual] = X_T
  all_X_NT[individual] = X_NT

individuals = list(subjects.keys())
fig = plt.figure(figsize=(15,7))

for individual in individuals:
  plot_mean_amp_per_indiv(all_X_T[individual], node_label='PO8', individual=individual, target=True)

individuals = list(subjects.keys())
fig = plt.figure(figsize=(15,7))

for individual in individuals:
  plot_mean_amp_per_indiv(all_X_NT[individual], node_label='PO8', individual=individual, target=False)

1.2.2 Spatial Information¶

Lastly I will see how it is all over the brain.

def plot_EEG_cmap_indiv(info, X, time, subplot, target=True):
  pos_list = dict()

  for i in range(1,len(info)+1):
    theta = np.radians(info.loc[i,'theta']+90)
    x = np.cos(theta)*info.loc[i,'radius']
    Y = np.sin(theta)*info.loc[i,'radius']
    pos_list[i]=(x,Y)

  node_labels = {i:info.loc[i,'class'].replace(' ','') for i in range(1,len(info)+1)}

  if (target):
    label = 'Target'
  else :
    label = 'Non-Target'
  
  plt.subplot(subplot)
  #plt.text(x=0, y=0.33, s='Nasion', horizontalalignment='c£enter');
  plt.text(x=0, y=-0.58, s='Inion', horizontalalignment='center', fontsize=13);


  g = nx.Graph()
  g.add_nodes_from(pos_list.keys())
  for i in pos_list.keys():
      g.nodes[i]['amp'] = np.mean(X[:, time, i-1])

  colors = list(nx.get_node_attributes(g,'amp').values())

  nc = nx.draw_networkx_nodes(g, pos_list, labels = node_labels, font_color= 'w', node_color=colors, node_size = 600, cmap=plt.cm.jet)
  nx.draw_networkx_labels(g, pos_list, labels = node_labels, font_color= 'w', font_size=10, font_weight= 'bold')

  cbar = plt.colorbar(nc)
  cbar.set_label('Amplitude (V)', rotation=270, fontsize=10, labelpad=15)
  plt.axis('off')
  plt.title('Average EEG result for one individual sequences at {}ms labeled as {} \n\n Nasion'.format(time, label))

fig = plt.figure(figsize=(20,7))

plot_EEG_cmap_indiv(info, all_X_NT[27], 35, 121, target=False)
plot_EEG_cmap_indiv(info, all_X_T[27], 35, 122, target=True)

plt.show()

fig = plt.figure(figsize=(20,7))

plot_EEG_cmap_indiv(info, all_X_NT[27], 250, 121, target=False)
plot_EEG_cmap_indiv(info, all_X_T[27], 250, 122, target=True)

plt.show()

1.2.3 Spectral Information¶

from scipy.signal import butter, lfilter, lfilter_zi

def butter_bandpass(lowcut, highcut, fs, order=5):
    nyq = 0.5 * fs
    low = lowcut / nyq
    high = highcut / nyq
    b, a = butter(order, [low, high], btype='band')
    return b, a


def butter_bandpass_filter(data, lowcut, highcut, fs, order=5):
    b, a = butter_bandpass(lowcut, highcut, fs, order=order)
    y = lfilter(b, a, data)
    return y

def get_label(node_label):
  map_node_labels = {1: 'Fz',
               2: 'F3',
               3: 'F4',
               4: 'Cz',
               5: 'C3',
               6: 'C4',
               7: 'Pz',
               8: 'P1',
               9: 'P2',
               10: 'P3',
               11: 'P4',
               12: 'POz',
               13: 'PO3',
               14: 'PO4',
               15: 'PO7',
               16: 'PO8',
               17: 'Oz',
               18: 'O1',
               19: 'O2'}
  
  inv_map_node_labels = {v: k for k, v in map_node_labels.items()}
  channel_nb = inv_map_node_labels[node_label]
  return channel_nb

def get_str_label(node_label):
  map_node_labels = {1: 'Fz',
               2: 'F3',
               3: 'F4',
               4: 'Cz',
               5: 'C3',
               6: 'C4',
               7: 'Pz',
               8: 'P1',
               9: 'P2',
               10: 'P3',
               11: 'P4',
               12: 'POz',
               13: 'PO3',
               14: 'PO4',
               15: 'PO7',
               16: 'PO8',
               17: 'Oz',
               18: 'O1',
               19: 'O2'}
  
  channel_nb = map_node_labels[node_label]
  return channel_nb

lowcut = 13
highcut = 30
orders = 3
node_label = get_label('POz')

yf = butter_bandpass_filter(X[301,:,node_label], lowcut, highcut, 1000, order=orders)

fig = plt.figure(figsize=(25,7))

plt.subplot(131)
plt.plot(X[301,:,node_label])
plt.title('Raw EEG at POz')
plt.xlabel('Time (ms)')
plt.ylabel('Amplitude (V)')

plt.subplot(132)
plt.plot(yf)
plt.title('Bandpass filter EEG at POz')
plt.xlabel('Time (ms)')
plt.ylabel('Amplitude (V)')

plt.subplot(133)
plt.plot(np.sqrt(yf**2))
plt.title('Power EEG at POz')
plt.xlabel('Time (ms)')
plt.ylabel('Amplitude (V)')

Text(0, 0.5, 'Amplitude (V)')

def mean_power(X, y, node_label, lowcut, highcut, orders):
  mean_powers = []
  labels = []
  epochs = []

  for epoch in range(len(X)):
    yf = butter_bandpass_filter(X[epoch,:,node_label], lowcut, highcut, 1000, order=orders)
    power = np.sqrt(yf**2)

    epochs.append(epoch)
    mean_powers.append(np.mean(power))
    labels.append(y[epoch])

  return pd.DataFrame({'epoch': epochs, 'mean power {}'.format(get_str_label(node_label)): mean_powers, 'label': labels})

lowcut = 13
highcut = 30
orders = 3
node_label = get_label('POz')

df_PO8 = mean_power(X, y, node_label, lowcut, highcut, orders)

node_label = get_label('Fz')

df_PO7 = mean_power(X, y, node_label, lowcut, highcut, orders)

df_PO7_8 = pd.merge(df_PO7.drop(['label'], axis=1), df_PO8, on='epoch', how='inner')
df_PO7_8

colors = ['red','green']
import matplotlib
#df_PO7_8.plot.scatter(x='mean power POz', y='mean power Fz', c='label', cmap=matplotlib.colors.ListedColormap(colors));

plt.scatter(df_PO7_8['mean power POz'].to_list(),
            df_PO7_8['mean power Fz'].to_list(),
            c=df_PO7_8['label'].to_list(),
            cmap=matplotlib.colors.ListedColormap(colors))
plt.title('Scattered points of mean power along POz and Fz')
plt.xlabel('Mean power POz')
plt.ylabel('Mean power Fz')

Text(0, 0.5, 'Mean power Fz')

2. Model selection¶

2.1. EEGNetv4¶

def cv_per_subj_val(x,y,params,path_to_subj, test_on_last_block=False, plot_fold_history=False):
    model_path = os.path.join(path_to_subj,'checkpoints')
    best_val_epochs = []
    best_val_aucs = []
    folds = 4  # To preserve split as 0.6 0.2 0.2
    x_tr,y_tr,x_tst,y_tst = separte_last_block(x,y,test_size=0.2)

    cv = StratifiedKFold(n_splits=folds, shuffle=True)
    cv_splits = list(cv.split(x_tr, y_tr))
    for fold, (train_idx, val_idx) in enumerate(cv_splits):
        fold_model_path = os.path.join(model_path, '%d' % fold)
        os.makedirs(fold_model_path)
        x_tr_fold, y_tr_fold = x_tr[train_idx], y_tr[train_idx]
        x_val_fold, y_val_fold = x_tr[val_idx], y_tr[val_idx]
        val_history, fold_model = train_model_eegnet(x_tr_fold,y_tr_fold,params,(x_val_fold,y_val_fold),epochs=25,
                                                     batch_size=32, shuffle=True,
                                                     model_path=os.path.join(fold_model_path,'model{}'.format(fold)))
        best_val_epochs.append(np.argmax(val_history['val_auc']) + 1)  # epochs count from 1 (not from 0)
        best_val_aucs.append(np.max(val_history['val_auc']))
        if plot_fold_history:
            single_auc_loging(val_history, 'fold %d' % fold, fold_model_path)
    _ , final_model = train_model_eegnet(x_tr, y_tr, params, epochs=int(np.mean(best_val_epochs)),
                                                       validation_data=(x_tst, y_tst), batch_size=32, shuffle=True,
                                                       model_path=os.path.join(path_to_subj,'naive_model'))

    single_auc_loging(val_history, 'val history', path_to_save=path_to_subj)
    with codecs.open('%s/res.txt' % path_to_subj, 'w', encoding='utf8') as f:
        f.write(u'Val auc %.02f±%.02f\n' % (np.mean(best_val_aucs),np.std(best_val_aucs)))
        f.write('Test auc naive %.02f\n' % (val_history['val_auc'][-1]))

    return {'val_auc':val_history['val_auc'][-1]}, final_model

def cv_per_subj_test(x,y,params,path_to_subj, test_on_last_block=False, plot_fold_history=False):
    model_path = os.path.join(path_to_subj,'checkpoints')
    best_val_epochs = []
    best_val_aucs = []
    folds = 4  # To preserve split as 0.6 0.2 0.2
    if test_on_last_block:
        x_tr,y_tr,x_tst,y_tst = separte_last_block(x,y,test_size=0.2)

    cv = StratifiedKFold(n_splits=folds, shuffle=True)
    cv_splits = list(cv.split(x_tr, y_tr))
    for fold, (train_idx, val_idx) in enumerate(cv_splits):
        fold_model_path = os.path.join(model_path, '%d' % fold)
        os.makedirs(fold_model_path)
        x_tr_fold, y_tr_fold = x_tr[train_idx], y_tr[train_idx]
        x_val_fold, y_val_fold = x_tr[val_idx], y_tr[val_idx]
        val_history, fold_model = train_model_eegnet(x_tr_fold,y_tr_fold,params,(x_val_fold,y_val_fold),epochs=200,
                                                     batch_size=32, shuffle=True,
                                                     model_path=os.path.join(fold_model_path,'model{}'.format(fold)))
        best_val_epochs.append(np.argmax(val_history['val_auc']) + 1)  # epochs count from 1 (not from 0)
        best_val_aucs.append(np.max(val_history['val_auc']))
        if plot_fold_history:
            single_auc_loging(val_history, 'fold %d' % fold, fold_model_path)

    if test_on_last_block:
        test_history, final_model = train_model_eegnet(x_tr, y_tr, params, epochs=int(np.mean(best_val_epochs)),
                                                       validation_data=(x_tst, y_tst), batch_size=32, shuffle=True,
                                                       model_path=os.path.join(path_to_subj,'naive_model'))

    single_auc_loging(test_history, 'test_history', path_to_save=path_to_subj)
    with codecs.open('%s/res.txt' % path_to_subj, 'w', encoding='utf8') as f:
        f.write(u'Val auc %.02f±%.02f\n' % (np.mean(best_val_aucs),np.std(best_val_aucs)))
        f.write('Test auc naive %.02f\n' % (test_history['val_auc'][-1]))

    return {'val_auc':test_history['val_auc'][-1]}, final_model

import sys
sys.path.append('/content/drive/My Drive/AML/Chal3 - EEG Classification/Code/')
from utils import single_auc_loging
from utils import prepare_dirs,write_results_table, separte_last_block
from model_torch import train_model_eegnet

from time import time

experiment_res_dir = './res/'
subjs_test_stats = {}
start = time()
for train_subject in all_subjects:
    path_to_subj = prepare_dirs(experiment_res_dir, train_subject)
    x = subjects[train_subject][0]
    x = x.transpose(0, 2, 1)[:, np.newaxis, :, :]
    y=subjects[train_subject][1]
    test_stats, model = cv_per_subj_val(x, y, params, path_to_subj,test_on_last_block=False, plot_fold_history=True)
    subjs_test_stats[train_subject] = test_stats
end = time()
running_time = end - start

Epoch 0: train loss 0.6812723619597298
Epoch 0: val loss 0.6728879809379578

Epoch 1: train loss 0.6377026012965611
Epoch 1: val loss 0.6454147100448608

Epoch 2: train loss 0.6145414454596383
Epoch 2: val loss 0.6278131604194641

Epoch 3: train loss 0.6095997946602958
Epoch 3: val loss 0.6127532720565796

Epoch 4: train loss 0.6044847667217255
Epoch 4: val loss 0.6012178063392639

Epoch 5: train loss 0.6174809421811785
Epoch 5: val loss 0.5939245223999023

Epoch 6: train loss 0.5477213902132851
Epoch 6: val loss 0.5890060067176819

Epoch 7: train loss 0.5122129150799343
Epoch 7: val loss 0.5850449800491333

Epoch 8: train loss 0.5799910170691354
Epoch 8: val loss 0.5824064016342163

Epoch 9: train loss 0.5593208244868687
Epoch 9: val loss 0.5811654925346375

Epoch 10: train loss 0.5145556075232369
Epoch 10: val loss 0.5823761820793152

Epoch 11: train loss 0.5030832077775683
Epoch 11: val loss 0.5856128334999084

Epoch 12: train loss 0.4884253016539982
Epoch 12: val loss 0.5878288745880127

Epoch 13: train loss 0.46808813725199017
Epoch 13: val loss 0.5885663628578186

Epoch 14: train loss 0.4912246423108237
Epoch 14: val loss 0.587224006652832

Epoch 15: train loss 0.4604695737361908
Epoch 15: val loss 0.5879901647567749

Epoch 16: train loss 0.4975290426186153
Epoch 16: val loss 0.5820066332817078

Epoch 17: train loss 0.4763775851045336
Epoch 17: val loss 0.58408522605896

Epoch 18: train loss 0.4826668756348746
Epoch 18: val loss 0.5875061750411987

Epoch 19: train loss 0.5079632529190609
Epoch 19: val loss 0.5872671008110046

Epoch 20: train loss 0.5433941398348127
Epoch 20: val loss 0.5911863446235657

Epoch 21: train loss 0.45587277838162016
Epoch 21: val loss 0.5932555198669434

Epoch 22: train loss 0.4961792060307094
Epoch 22: val loss 0.5976765155792236

Epoch 23: train loss 0.5032381074769157
Epoch 23: val loss 0.6026989221572876

Epoch 24: train loss 0.5286267995834351
Epoch 24: val loss 0.5972000956535339

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.6776858227593558
Epoch 0: val loss 0.6530750393867493

Epoch 1: train loss 0.6062617812837873
Epoch 1: val loss 0.6309542059898376

Epoch 2: train loss 0.5537767665726798
Epoch 2: val loss 0.615822970867157

Epoch 3: train loss 0.5693877594811576
Epoch 3: val loss 0.6105405688285828

Epoch 4: train loss 0.5430010386875698
Epoch 4: val loss 0.6231949329376221

Epoch 5: train loss 0.5091979844229562
Epoch 5: val loss 0.6116743683815002

Epoch 6: train loss 0.5228845136506217
Epoch 6: val loss 0.6180281043052673

Epoch 7: train loss 0.49188846349716187
Epoch 7: val loss 0.6046333909034729

Epoch 8: train loss 0.4900141443525042
Epoch 8: val loss 0.599396288394928

Epoch 9: train loss 0.4806723807539259
Epoch 9: val loss 0.6020609140396118

Epoch 10: train loss 0.49785180177007404
Epoch 10: val loss 0.6088827252388

Epoch 11: train loss 0.5361165574618748
Epoch 11: val loss 0.6208125948905945

Epoch 12: train loss 0.5313577694552285
Epoch 12: val loss 0.6397662162780762

Epoch 13: train loss 0.47643967611449106
Epoch 13: val loss 0.6265682578086853

Epoch 14: train loss 0.47546193855149405
Epoch 14: val loss 0.6290757060050964

Epoch 15: train loss 0.4676474332809448
Epoch 15: val loss 0.6332340240478516

Epoch 16: train loss 0.4904316876615797
Epoch 16: val loss 0.6315407156944275

Epoch 17: train loss 0.45331049816949026
Epoch 17: val loss 0.6325690746307373

Epoch 18: train loss 0.49719343866620747
Epoch 18: val loss 0.6365164518356323

Epoch 19: train loss 0.47648257868630545
Epoch 19: val loss 0.6399645209312439

Epoch 20: train loss 0.48656447870390757
Epoch 20: val loss 0.646192193031311

Epoch 21: train loss 0.4455475764615195
Epoch 21: val loss 0.6596159338951111

Epoch 22: train loss 0.46718467133385794
Epoch 22: val loss 0.6657642126083374

Epoch 23: train loss 0.46082304630960735
Epoch 23: val loss 0.6633044481277466

Epoch 24: train loss 0.43993800452777315
Epoch 24: val loss 0.6596046686172485

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.6882990002632141
Epoch 0: val loss 0.6562791466712952

Epoch 1: train loss 0.6443651999746051
Epoch 1: val loss 0.6167645454406738

Epoch 2: train loss 0.6067640951701573
Epoch 2: val loss 0.5913919806480408

Epoch 3: train loss 0.5804405042103359
Epoch 3: val loss 0.5766610503196716

Epoch 4: train loss 0.6444420644215175
Epoch 4: val loss 0.5682393908500671

Epoch 5: train loss 0.5476283388478416
Epoch 5: val loss 0.5616909861564636

Epoch 6: train loss 0.580171252999987
Epoch 6: val loss 0.558915376663208

Epoch 7: train loss 0.5355492021356311
Epoch 7: val loss 0.5606633424758911

Epoch 8: train loss 0.5685944727488926
Epoch 8: val loss 0.5591132640838623

Epoch 9: train loss 0.5124402642250061
Epoch 9: val loss 0.5568347573280334

Epoch 10: train loss 0.5458018268857684
Epoch 10: val loss 0.5575602054595947

Epoch 11: train loss 0.6237334949629647
Epoch 11: val loss 0.5614291429519653

Epoch 12: train loss 0.4930976331233978
Epoch 12: val loss 0.5684775114059448

Epoch 13: train loss 0.5011431447097233
Epoch 13: val loss 0.5710557103157043

Epoch 14: train loss 0.5276197195053101
Epoch 14: val loss 0.5707655549049377

Epoch 15: train loss 0.5075239326272692
Epoch 15: val loss 0.5715076923370361

Epoch 16: train loss 0.5139421692916325
Epoch 16: val loss 0.5734698176383972

Epoch 17: train loss 0.4720810651779175
Epoch 17: val loss 0.5736613869667053

Epoch 18: train loss 0.4916122555732727
Epoch 18: val loss 0.5703460574150085

Epoch 19: train loss 0.4702781481402261
Epoch 19: val loss 0.5736677646636963

Epoch 20: train loss 0.5062455620084491
Epoch 20: val loss 0.579793393611908

Epoch 21: train loss 0.4615068903991154
Epoch 21: val loss 0.5935271382331848

Epoch 22: train loss 0.4402706410203661
Epoch 22: val loss 0.5970081090927124

Epoch 23: train loss 0.5399354909147535
Epoch 23: val loss 0.5906126499176025

Epoch 24: train loss 0.43179048384938923
Epoch 24: val loss 0.5938133001327515

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.6913323998451233
Epoch 0: val loss 0.62083500623703

Epoch 1: train loss 0.6320870092936924
Epoch 1: val loss 0.5890552997589111

Epoch 2: train loss 0.5788610279560089
Epoch 2: val loss 0.5770668387413025

Epoch 3: train loss 0.5745051077433995
Epoch 3: val loss 0.5720250010490417

Epoch 4: train loss 0.5812374310834068
Epoch 4: val loss 0.5769871473312378

Epoch 5: train loss 0.5441465292658124
Epoch 5: val loss 0.5886284112930298

Epoch 6: train loss 0.5370239359991891
Epoch 6: val loss 0.6001216173171997

Epoch 7: train loss 0.5265422207968575
Epoch 7: val loss 0.6069337129592896

Epoch 8: train loss 0.526741508926664
Epoch 8: val loss 0.6103532314300537

Epoch 9: train loss 0.5168006292411259
Epoch 9: val loss 0.6090015172958374

Epoch 10: train loss 0.5260582864284515
Epoch 10: val loss 0.6155492067337036

Epoch 11: train loss 0.5119486408574241
Epoch 11: val loss 0.6196824908256531

Epoch 12: train loss 0.531713822058269
Epoch 12: val loss 0.6226046681404114

Epoch 13: train loss 0.5069538695471627
Epoch 13: val loss 0.6118451356887817

Epoch 14: train loss 0.4961351156234741
Epoch 14: val loss 0.6294457912445068

Epoch 15: train loss 0.508637011051178
Epoch 15: val loss 0.6328803896903992

Epoch 16: train loss 0.504221362727029
Epoch 16: val loss 0.6251507997512817

Epoch 17: train loss 0.503167097057615
Epoch 17: val loss 0.6228746771812439

Epoch 18: train loss 0.5124055785792214
Epoch 18: val loss 0.6482216715812683

Epoch 19: train loss 0.4720651720251356
Epoch 19: val loss 0.6396117210388184

Epoch 20: train loss 0.46857720613479614
Epoch 20: val loss 0.6263028383255005

Epoch 21: train loss 0.4805868864059448
Epoch 21: val loss 0.6321181654930115

Epoch 22: train loss 0.45765596628189087
Epoch 22: val loss 0.6412585377693176

Epoch 23: train loss 0.43052880678858074
Epoch 23: val loss 0.6572952270507812

Epoch 24: train loss 0.45988430295671734
Epoch 24: val loss 0.6639794707298279

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.6425549189249674
Epoch 0: val loss 0.6175201535224915

Epoch 1: train loss 0.5756158464484744
Epoch 1: val loss 0.5860295295715332

Epoch 2: train loss 0.5674390196800232
Epoch 2: val loss 0.5875415802001953

Epoch 3: train loss 0.565978639655643
Epoch 3: val loss 0.5929503440856934

Epoch 4: train loss 0.5391473604573144
Epoch 4: val loss 0.6058016419410706

Epoch 5: train loss 0.5410087704658508
Epoch 5: val loss 0.6292356848716736

Epoch 6: train loss 0.5571317705843184
Epoch 6: val loss 0.6329464316368103

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

subjs_test_stats

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-1-db57cf2a24e2> in <module>()
----> 1 subjs_test_stats

NameError: name 'subjs_test_stats' is not defined

running_time

157.6341691017151

f1_list = [int(x) for x in np.linspace(start=9, stop= 15, num = 7)]
D_list = [x for x in range (1,6)]
search_grid = {'resample_to': [369],
                 'D': D_list,
                 'F1': f1_list,
                 'dropoutRate1': [0.52],
                 'dropoutRate2': [0.36],
                 'lr': [0.00066],
                 'norm_rate': [0.275]
                 }

from sklearn.model_selection import ParameterGrid
params_list = list(ParameterGrid(search_grid))
experiment_res_dir = './res/'
subj_tests_stats_all = {}
params_index = 0
start_grid = time()
for params in params_list:
  subjs_test_stats_grid = {}
  for train_subject in all_subjects:
    path_to_subj = prepare_dirs(experiment_res_dir, train_subject)
    x = subjects[train_subject][0]
    x = x.transpose(0, 2, 1)[:, np.newaxis, :, :]
    y=subjects[train_subject][1]
    test_stats, model = cv_per_subj_val(x, y, params, path_to_subj,test_on_last_block=True, plot_fold_history=True)
    subjs_test_stats_grid[train_subject] = test_stats
  subj_tests_stats_all[params_index] = subjs_test_stats_grid
  params_index += 1
  print (params_index)
  end_grid = time()
  running_time_grid = end_grid - start_grid

Epoch 0: train loss 0.6715642980166844
Epoch 0: val loss 0.6928219199180603

Epoch 1: train loss 0.6705542547362191
Epoch 1: val loss 0.674240231513977

Epoch 2: train loss 0.6701717291559491
Epoch 2: val loss 0.6543846130371094

Epoch 3: train loss 0.6463924050331116
Epoch 3: val loss 0.6314519643783569

Epoch 4: train loss 0.6079713446753365
Epoch 4: val loss 0.6164222359657288

Epoch 5: train loss 0.6063820804868426
Epoch 5: val loss 0.5973367691040039

Epoch 6: train loss 0.6219964368002755
Epoch 6: val loss 0.5901148319244385

Epoch 7: train loss 0.612184464931488
Epoch 7: val loss 0.5815602540969849

Epoch 8: train loss 0.6180401529584613
Epoch 8: val loss 0.5710257291793823

Epoch 9: train loss 0.5997190220015389
Epoch 9: val loss 0.585863471031189

Epoch 10: train loss 0.5669413123811994
Epoch 10: val loss 0.5796632170677185

Epoch 11: train loss 0.5837877988815308
Epoch 11: val loss 0.5554513335227966

Epoch 12: train loss 0.5423993425709861
Epoch 12: val loss 0.5486589074134827

Epoch 13: train loss 0.5722302113260541
Epoch 13: val loss 0.552062451839447

Epoch 14: train loss 0.5416759635720935
Epoch 14: val loss 0.5572354197502136

Epoch 15: train loss 0.5586699885981423
Epoch 15: val loss 0.5664275884628296

Epoch 16: train loss 0.5445415803364345
Epoch 16: val loss 0.5743698477745056

Epoch 17: train loss 0.5867940187454224
Epoch 17: val loss 0.5723559856414795

Epoch 18: train loss 0.5209656315190452
Epoch 18: val loss 0.5712056159973145

Epoch 19: train loss 0.5744568364960807
Epoch 19: val loss 0.5800946950912476

Epoch 20: train loss 0.5806652435234615
Epoch 20: val loss 0.5950448513031006

Epoch 21: train loss 0.5984036113534655
Epoch 21: val loss 0.5895622372627258

Epoch 22: train loss 0.5643349758216313
Epoch 22: val loss 0.5708938241004944

Epoch 23: train loss 0.5647242196968624
Epoch 23: val loss 0.5566393733024597

Epoch 24: train loss 0.5624855416161674
Epoch 24: val loss 0.54775470495224

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.7161681226321629
Epoch 0: val loss 0.701287567615509

Epoch 1: train loss 0.7012851153101239
Epoch 1: val loss 0.6883065104484558

Epoch 2: train loss 0.6963419999395098
Epoch 2: val loss 0.6751147508621216

Epoch 3: train loss 0.6590280703135899
Epoch 3: val loss 0.659736156463623

Epoch 4: train loss 0.6786387733050755
Epoch 4: val loss 0.645875871181488

Epoch 5: train loss 0.637402321611132
Epoch 5: val loss 0.6366515755653381

Epoch 6: train loss 0.6350091780935015
Epoch 6: val loss 0.6267271041870117

Epoch 7: train loss 0.6001926915986198
Epoch 7: val loss 0.6171947121620178

Epoch 8: train loss 0.5875019942011152
Epoch 8: val loss 0.6131923794746399

Epoch 9: train loss 0.612590491771698
Epoch 9: val loss 0.6103071570396423

Epoch 10: train loss 0.6170480293887002
Epoch 10: val loss 0.6085920333862305

Epoch 11: train loss 0.5852334499359131
Epoch 11: val loss 0.603973925113678

Epoch 12: train loss 0.5562644132546016
Epoch 12: val loss 0.6019946932792664

Epoch 13: train loss 0.6018508757863726
Epoch 13: val loss 0.5991787910461426

Epoch 14: train loss 0.6040815157549722
Epoch 14: val loss 0.5990949869155884

Epoch 15: train loss 0.5785707405635289
Epoch 15: val loss 0.5969551801681519

Epoch 16: train loss 0.5625947628702436
Epoch 16: val loss 0.5989424586296082

Epoch 17: train loss 0.5621133531842913
Epoch 17: val loss 0.5960992574691772

Epoch 18: train loss 0.5805429645947048
Epoch 18: val loss 0.6024590730667114

Epoch 19: train loss 0.560452265398843
Epoch 19: val loss 0.6080248951911926

Epoch 20: train loss 0.544825839144843
Epoch 20: val loss 0.616115927696228

Epoch 21: train loss 0.5214542703969138
Epoch 21: val loss 0.619403600692749

Epoch 22: train loss 0.5621991583279201
Epoch 22: val loss 0.6168621182441711

Epoch 23: train loss 0.6451768279075623
Epoch 23: val loss 0.6068918108940125

Epoch 24: train loss 0.5869414934090206
Epoch 24: val loss 0.6030218601226807

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.7215496301651001
Epoch 0: val loss 0.7089601755142212

Epoch 1: train loss 0.6889938030924115
Epoch 1: val loss 0.6950243711471558

Epoch 2: train loss 0.6669026272637504
Epoch 2: val loss 0.6818302869796753

Epoch 3: train loss 0.6439823167664664
Epoch 3: val loss 0.6662390232086182

Epoch 4: train loss 0.6853544797216143
Epoch 4: val loss 0.6488272547721863

Epoch 5: train loss 0.645276061126164
Epoch 5: val loss 0.6374397873878479

Epoch 6: train loss 0.6165140015738351
Epoch 6: val loss 0.6242775321006775

Epoch 7: train loss 0.6047137039048331
Epoch 7: val loss 0.6138905882835388

Epoch 8: train loss 0.5915160093988691
Epoch 8: val loss 0.6019322276115417

Epoch 9: train loss 0.6044481907572065
Epoch 9: val loss 0.5974265933036804

Epoch 10: train loss 0.5470473510878426
Epoch 10: val loss 0.595832347869873

Epoch 11: train loss 0.5984078177383968
Epoch 11: val loss 0.5947847962379456

Epoch 12: train loss 0.5568753225462777
Epoch 12: val loss 0.5959655046463013

Epoch 13: train loss 0.5533498866217477
Epoch 13: val loss 0.5979648232460022

Epoch 14: train loss 0.5950216778687069
Epoch 14: val loss 0.5995094776153564

Epoch 15: train loss 0.5641294462340218
Epoch 15: val loss 0.6007103323936462

Epoch 16: train loss 0.5351525800568717
Epoch 16: val loss 0.6006554365158081

Epoch 17: train loss 0.5362184984343392
Epoch 17: val loss 0.6021566390991211

Epoch 18: train loss 0.5406815026487622
Epoch 18: val loss 0.6023834943771362

Epoch 19: train loss 0.5185949546950204
Epoch 19: val loss 0.6032992601394653

Epoch 20: train loss 0.5124561062880925
Epoch 20: val loss 0.6078791618347168

Epoch 21: train loss 0.5246328370911735
Epoch 21: val loss 0.613617479801178

Epoch 22: train loss 0.5187411436012813
Epoch 22: val loss 0.6131135821342468

Epoch 23: train loss 0.512135135275977
Epoch 23: val loss 0.6137229800224304

Epoch 24: train loss 0.6279764771461487
Epoch 24: val loss 0.6156291365623474

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.6925899641854423
Epoch 0: val loss 0.6859037280082703

Epoch 1: train loss 0.6580914429255894
Epoch 1: val loss 0.6756787896156311

Epoch 2: train loss 0.646507740020752
Epoch 2: val loss 0.6652714610099792

Epoch 3: train loss 0.6597482647214618
Epoch 3: val loss 0.6534170508384705

Epoch 4: train loss 0.6341134990964618
Epoch 4: val loss 0.64519202709198

Epoch 5: train loss 0.6013632331575666
Epoch 5: val loss 0.6358400583267212

Epoch 6: train loss 0.6199805395943778
Epoch 6: val loss 0.6275656223297119

Epoch 7: train loss 0.591161813054766
Epoch 7: val loss 0.6231477856636047

Epoch 8: train loss 0.5995940566062927
Epoch 8: val loss 0.61635422706604

Epoch 9: train loss 0.5762082806655339
Epoch 9: val loss 0.6111354827880859

Epoch 10: train loss 0.5945536323956081
Epoch 10: val loss 0.6073857545852661

Epoch 11: train loss 0.5592281350067684
Epoch 11: val loss 0.6033686399459839

Epoch 12: train loss 0.5559146063668388
Epoch 12: val loss 0.6027696132659912

Epoch 13: train loss 0.5937318333557674
Epoch 13: val loss 0.6032614707946777

Epoch 14: train loss 0.5694811940193176
Epoch 14: val loss 0.6013914942741394

Epoch 15: train loss 0.6208419586930957
Epoch 15: val loss 0.5983809232711792

Epoch 16: train loss 0.5845959527151925
Epoch 16: val loss 0.597998857498169

Epoch 17: train loss 0.5428755581378937
Epoch 17: val loss 0.5965712070465088

Epoch 18: train loss 0.5474284248692649
Epoch 18: val loss 0.5955135226249695

Epoch 19: train loss 0.5488146458353315
Epoch 19: val loss 0.5914111137390137

Epoch 20: train loss 0.5650747460978371
Epoch 20: val loss 0.5933457612991333

Epoch 21: train loss 0.5513074483190264
Epoch 21: val loss 0.591745138168335

Epoch 22: train loss 0.5236807380403791
Epoch 22: val loss 0.5881955027580261

Epoch 23: train loss 0.523570933512279
Epoch 23: val loss 0.5883394479751587

Epoch 24: train loss 0.5472377070358821
Epoch 24: val loss 0.5878168344497681

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.6846602890226576
Epoch 0: val loss 0.6260190606117249

Epoch 1: train loss 0.6248641146553887
Epoch 1: val loss 0.6180055141448975

Epoch 2: train loss 0.619761221938663
Epoch 2: val loss 0.6173037886619568

Epoch 3: train loss 0.5933655301729838
Epoch 3: val loss 0.618974506855011

Epoch 4: train loss 0.5973031487729814
Epoch 4: val loss 0.6163362264633179

Epoch 5: train loss 0.573019865486357
Epoch 5: val loss 0.6145980358123779

Epoch 6: train loss 0.5635498099856906
Epoch 6: val loss 0.6113642454147339

Epoch 7: train loss 0.5972888900174035
Epoch 7: val loss 0.6087958812713623

Epoch 8: train loss 0.5769046280119154
Epoch 8: val loss 0.6080853343009949

Epoch 9: train loss 0.5998044775591956
Epoch 9: val loss 0.6068524122238159

Epoch 10: train loss 0.5841925773355696
Epoch 10: val loss 0.6130109429359436

Epoch 11: train loss 0.5784846014446683
Epoch 11: val loss 0.6152065396308899

Epoch 12: train loss 0.5522556636068556
Epoch 12: val loss 0.6275192499160767

Epoch 13: train loss 0.5563968221346537
Epoch 13: val loss 0.629492461681366

Epoch 14: train loss 0.5625432564152612
Epoch 14: val loss 0.626389741897583

Epoch 15: train loss 0.5537484751807319
Epoch 15: val loss 0.6166159510612488

Epoch 16: train loss 0.54015937116411
Epoch 16: val loss 0.6210877299308777

Epoch 17: train loss 0.540546284781562
Epoch 17: val loss 0.627122700214386

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

%matplotlib inline

fig = plt.figure(figsize=(40,20))
index=[k for k in range (14)]
index.remove(6)
for k in index:
  parameters = []
  auc_values = []
  for i in range(len(subj_tests_stats_all)):
    parameters.append(i)
    auc_values.append(list((subj_tests_stats_all[i][25+k].values())).pop())
  plt.subplot(3,5,k+1)
  sns.barplot(parameters, auc_values)
  plt.title('AUC for patient number '+ str(25+k))
  plt.xlabel('set of parameters')
  plt.ylabel('auc validation loss')
plt.show()

auc_values_mean = []
for i in range(len(subj_tests_stats_all)):
  for k in index:
    temp = []
    temp.append(list((subj_tests_stats_all[i][25+k].values())).pop())
  avg = sum(temp) / len(temp)
  auc_values_mean.append(avg)

fig = plt.figure(figsize=(6,5))
sns.barplot(parameters, auc_values_mean)
plt.title('AUC validation loss averaged over patients')
plt.xlabel('parameters set')
plt.ylabel('auc validation loss')
plt.show()

params_list[6]

{'D': 1,
 'F1': 15,
 'dropoutRate1': 0.52,
 'dropoutRate2': 0.36,
 'lr': 0.00066,
 'norm_rate': 0.275,
 'resample_to': 369}

from time import time

experiment_res_dir = './res/'
subjs_test_stats = {}
start = time()
for train_subject in all_subjects:
    path_to_subj = prepare_dirs(experiment_res_dir, train_subject)
    x = subjects[train_subject][0]
    x = x.transpose(0, 2, 1)[:, np.newaxis, :, :]
    y=subjects[train_subject][1]
    test_stats, model = cv_per_subj_test(x, y, params, path_to_subj,test_on_last_block=False, plot_fold_history=True)
    subjs_test_stats[train_subject] = test_stats
end = time()
running_time = end - start

Epoch 0: train loss 0.7075626424380711
Epoch 0: val loss 0.6693019270896912

Epoch 1: train loss 0.6550172482218061
Epoch 1: val loss 0.6492401361465454

Epoch 2: train loss 0.626555894102369
Epoch 2: val loss 0.6292676329612732

Epoch 3: train loss 0.5704562153135028
Epoch 3: val loss 0.6159329414367676

Epoch 4: train loss 0.5912736824580601
Epoch 4: val loss 0.6122496724128723

Epoch 5: train loss 0.6688347969736371
Epoch 5: val loss 0.6068553328514099

Epoch 6: train loss 0.5628789067268372
Epoch 6: val loss 0.5980691909790039

Epoch 7: train loss 0.5136201637131828
Epoch 7: val loss 0.5973156094551086

Epoch 8: train loss 0.5590437224933079
Epoch 8: val loss 0.5966963768005371

Epoch 9: train loss 0.5242980165140969
Epoch 9: val loss 0.5982412695884705

Epoch 10: train loss 0.49803513288497925
Epoch 10: val loss 0.6022525429725647

Epoch 11: train loss 0.5224203509943826
Epoch 11: val loss 0.6041831374168396

Epoch 12: train loss 0.5636172720364162
Epoch 12: val loss 0.6078053116798401

Epoch 13: train loss 0.5023358379091535
Epoch 13: val loss 0.6110040545463562

Epoch 14: train loss 0.5605086939675468
Epoch 14: val loss 0.6142551898956299

Epoch 15: train loss 0.5044092621122088
Epoch 15: val loss 0.6164275407791138

Epoch 16: train loss 0.473884825195585
Epoch 16: val loss 0.6188696026802063

Epoch 17: train loss 0.5212748391287667
Epoch 17: val loss 0.6202488541603088

Epoch 18: train loss 0.5235577779156821
Epoch 18: val loss 0.61970454454422

Epoch 19: train loss 0.4778010674885341
Epoch 19: val loss 0.625593364238739

Epoch 20: train loss 0.48105837191854206
Epoch 20: val loss 0.6330726742744446

Epoch 21: train loss 0.47507867217063904
Epoch 21: val loss 0.6333102583885193

Epoch 22: train loss 0.5113307109900883
Epoch 22: val loss 0.6330316066741943

Epoch 23: train loss 0.49526710169655935
Epoch 23: val loss 0.6438152194023132

Epoch 24: train loss 0.4738272896834782
Epoch 24: val loss 0.6533827185630798

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.7000317062650409
Epoch 0: val loss 0.6655334234237671

Epoch 1: train loss 0.6445214322635106
Epoch 1: val loss 0.6374626159667969

Epoch 2: train loss 0.5889539846352169
Epoch 2: val loss 0.6129174828529358

Epoch 3: train loss 0.5625937410763332
Epoch 3: val loss 0.593253493309021

Epoch 4: train loss 0.5452094546386174
Epoch 4: val loss 0.5790622234344482

Epoch 5: train loss 0.49977693387440275
Epoch 5: val loss 0.5732026100158691

Epoch 6: train loss 0.5086701810359955
Epoch 6: val loss 0.5724141597747803

Epoch 7: train loss 0.4995333807809012
Epoch 7: val loss 0.5748735070228577

Epoch 8: train loss 0.514498553105763
Epoch 8: val loss 0.5752103924751282

Epoch 9: train loss 0.4586645875658308
Epoch 9: val loss 0.5722005367279053

Epoch 10: train loss 0.5242529852049691
Epoch 10: val loss 0.5745800137519836

Epoch 11: train loss 0.5455842529024396
Epoch 11: val loss 0.576036274433136

Epoch 12: train loss 0.4978488045079367
Epoch 12: val loss 0.5927244424819946

Epoch 13: train loss 0.47448123352868216
Epoch 13: val loss 0.5887723565101624

Epoch 14: train loss 0.4802858829498291
Epoch 14: val loss 0.5751505494117737

Epoch 15: train loss 0.45370841664927347
Epoch 15: val loss 0.576065182685852

Epoch 16: train loss 0.4560088515281677
Epoch 16: val loss 0.5787676572799683

Epoch 17: train loss 0.430768010871751
Epoch 17: val loss 0.5789093375205994

Epoch 18: train loss 0.4214467150824411
Epoch 18: val loss 0.5868712663650513

Epoch 19: train loss 0.4878957271575928
Epoch 19: val loss 0.5949043035507202

Epoch 20: train loss 0.42892466698374065
Epoch 20: val loss 0.6113376021385193

Epoch 21: train loss 0.5357583633490971
Epoch 21: val loss 0.6219988465309143

Epoch 22: train loss 0.4613423730645861
Epoch 22: val loss 0.6279308795928955

Epoch 23: train loss 0.45673652631895884
Epoch 23: val loss 0.6307450532913208

Epoch 24: train loss 0.45194070254053387
Epoch 24: val loss 0.6240681409835815

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.618015740598951
Epoch 0: val loss 0.6697432994842529

Epoch 1: train loss 0.6237448709351676
Epoch 1: val loss 0.6623064875602722

Epoch 2: train loss 0.5649449144090924
Epoch 2: val loss 0.6392761468887329

Epoch 3: train loss 0.5431296484810966
Epoch 3: val loss 0.6213893890380859

Epoch 4: train loss 0.5354042989867074
Epoch 4: val loss 0.6126813888549805

Epoch 5: train loss 0.5041447963033404
Epoch 5: val loss 0.6121792197227478

Epoch 6: train loss 0.5090656663690295
Epoch 6: val loss 0.603278636932373

Epoch 7: train loss 0.5298337170055935
Epoch 7: val loss 0.5944743752479553

Epoch 8: train loss 0.4955692248685019
Epoch 8: val loss 0.6073456406593323

Epoch 9: train loss 0.5202823621886117
Epoch 9: val loss 0.6237495541572571

Epoch 10: train loss 0.4992067813873291
Epoch 10: val loss 0.6034252047538757

Epoch 11: train loss 0.4853321313858032
Epoch 11: val loss 0.5901617407798767

Epoch 12: train loss 0.4682281230177198
Epoch 12: val loss 0.6080827116966248

Epoch 13: train loss 0.4798535704612732
Epoch 13: val loss 0.6555761694908142

Epoch 14: train loss 0.5066387483051845
Epoch 14: val loss 0.6675098538398743

Epoch 15: train loss 0.47630258543150766
Epoch 15: val loss 0.6350551843643188

Epoch 16: train loss 0.48655180420194355
Epoch 16: val loss 0.6444180607795715

Epoch 17: train loss 0.49895509651729036
Epoch 17: val loss 0.6714121699333191

Epoch 18: train loss 0.4217871427536011
Epoch 18: val loss 0.6634523272514343

Epoch 19: train loss 0.4995980092457363
Epoch 19: val loss 0.6324561834335327

Epoch 20: train loss 0.463951975107193
Epoch 20: val loss 0.6016619205474854

Epoch 21: train loss 0.4594044642789023
Epoch 21: val loss 0.6196800470352173

Epoch 22: train loss 0.4319181186812265
Epoch 22: val loss 0.6171748638153076

Epoch 23: train loss 0.4560444397585733
Epoch 23: val loss 0.6438695788383484

Epoch 24: train loss 0.4264556808131082
Epoch 24: val loss 0.6595199108123779

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

Epoch 0: train loss 0.6671404497964042
Epoch 0: val loss 0.6138673424720764

Epoch 1: train loss 0.6537669726780483
Epoch 1: val loss 0.5890045166015625

Epoch 2: train loss 0.6288341879844666
Epoch 2: val loss 0.5761247873306274

Epoch 3: train loss 0.628504331622805
Epoch 3: val loss 0.5710749626159668

Epoch 4: train loss 0.5687848925590515
Epoch 4: val loss 0.5700005888938904

Epoch 5: train loss 0.5395081724439349
Epoch 5: val loss 0.5675393342971802

Epoch 6: train loss 0.5511469926152911
Epoch 6: val loss 0.568970799446106

Epoch 7: train loss 0.5766363229070391
Epoch 7: val loss 0.5771312117576599

Epoch 8: train loss 0.5216581438268934
Epoch 8: val loss 0.5814802050590515

Epoch 9: train loss 0.5162404307297298
Epoch 9: val loss 0.5790902376174927

Epoch 10: train loss 0.5067295602389744
Epoch 10: val loss 0.5820879340171814

Epoch 11: train loss 0.5231956158365522
Epoch 11: val loss 0.5851179957389832

Epoch 12: train loss 0.49377842460359844
Epoch 12: val loss 0.5873914361000061

Epoch 13: train loss 0.48914723311151775
Epoch 13: val loss 0.5977297425270081

Epoch 14: train loss 0.47863893849509104
Epoch 14: val loss 0.6007556319236755

Epoch 15: train loss 0.47175793775490354
Epoch 15: val loss 0.5982775688171387

Epoch 16: train loss 0.4719164158616747
Epoch 16: val loss 0.607292890548706

Epoch 17: train loss 0.5546867975166866
Epoch 17: val loss 0.6177462935447693

Epoch 18: train loss 0.4815893386091505
Epoch 18: val loss 0.6168175935745239

Epoch 19: train loss 0.48954073446137564
Epoch 19: val loss 0.6137657761573792

Epoch 20: train loss 0.48395353981426786
Epoch 20: val loss 0.6306226849555969

Epoch 21: train loss 0.48115886534963337
Epoch 21: val loss 0.6364952921867371

Epoch 22: train loss 0.4542894661426544
Epoch 22: val loss 0.6425607204437256

Epoch 23: train loss 0.48467434304101126
Epoch 23: val loss 0.6413478851318359

Epoch 24: train loss 0.4449761722769056
Epoch 24: val loss 0.6295065879821777

/usr/local/lib/python3.6/dist-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)

---------------------------------------------------------------------------
UnboundLocalError                         Traceback (most recent call last)
<ipython-input-14-3f9847ebd7fb> in <module>()
      9     x = x.transpose(0, 2, 1)[:, np.newaxis, :, :]
     10     y=subjects[train_subject][1]
---> 11     test_stats, model = cv_per_subj_test(x, y, params, path_to_subj,test_on_last_block=False, plot_fold_history=True)
     12     subjs_test_stats[train_subject] = test_stats
     13 end = time()

<ipython-input-12-d3c791c1423e> in cv_per_subj_test(x, y, params, path_to_subj, test_on_last_block, plot_fold_history)
     26                                                        model_path=os.path.join(path_to_subj,'naive_model'))
     27 
---> 28     single_auc_loging(test_history, 'test_history', path_to_save=path_to_subj)
     29     with codecs.open('%s/res.txt' % path_to_subj, 'w', encoding='utf8') as f:
     30         f.write(u'Val auc %.02f±%.02f\n' % (np.mean(best_val_aucs),np.std(best_val_aucs)))

UnboundLocalError: local variable 'test_history' referenced before assignment

indexes = [25 + k for k in range(14)]
indexes.remove(31)
plt.figure(figsize=(30,15))
n=1
for i in indexes:
  plt.subplot(3,5,n)
  n+=1
  img = plt.imread('/content/res/'+str(i)+'/test_history.png')
  plt.imshow(img)
plt.show()

2.1. LDA Classifier¶

We decided to use Beta bandwidth [13-30Hz]

lowcut = 13
highcut = 30
orders = 3
node_label = get_label('PO8')

yf = butter_bandpass_filter(X[301,:,node_label], lowcut, highcut, 1000, order=orders)

fig = plt.figure(figsize=(25,7))

plt.subplot(131)
plt.plot(X[301,:,node_label])

plt.subplot(132)
plt.plot(yf)

plt.subplot(133)
plt.plot(np.sqrt(yf**2))

[<matplotlib.lines.Line2D at 0x7fbcc03b6828>]

np.mean(np.sqrt(yf**2))

1.4981284847282739

2.1.1 Spatial filtering¶

First we needed to do feature extraction. According to the exploration we did up above, we realized that it could be relevant to do spatial filetring using the CSP algorithm, Common Space Pattern, to identify wether it is Target or a Non-Target sequence.

from mne.decoding import CSP

# run CSP
csp = CSP(reg='auto')
csp.fit(X,y)

Computing rank from data with rank=None
    Using tolerance 2.8e+02 (2.2e-16 eps * 300 dim * 4.2e+15  max singular value)
    Estimated rank (mag): 170
    MAG: rank 170 computed from 300 data channels with 0 projectors
    Setting small MAG eigenvalues to zero (without PCA)
Reducing data rank from 300 -> 170
Estimating covariance using AUTO

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-78-72507668dc8a> in <module>()
      3 # run CSP
      4 csp = CSP(reg='auto')
----> 5 csp.fit(X,y)

/usr/local/lib/python3.6/dist-packages/mne/decoding/csp.py in fit(self, X, y)
    176                 cov = _regularized_covariance(
    177                     class_, reg=self.reg, method_params=self.cov_method_params,
--> 178                     rank=self.rank)
    179                 weight = sum(y == this_class)
    180             elif self.cov_est == "epoch":

/usr/local/lib/python3.6/dist-packages/mne/cov.py in _regularized_covariance(data, reg, method_params, info, rank)
   1672         info=info, cv=None, n_jobs=1, stop_early=True,
   1673         picks_list=picks_list, scalings=scalings,
-> 1674         rank=rank)[reg]['data']
   1675     return cov
   1676 

/usr/local/lib/python3.6/dist-packages/mne/cov.py in _compute_covariance_auto(data, method, info, method_params, cv, scalings, n_jobs, stop_early, picks_list, rank)
    974             name = method_.__name__ if callable(method_) else method_
    975             logger.info(msg % name.upper())
--> 976             mp = method_params[method_]
    977             _info = {}
    978 

KeyError: 'auto'

csp

CSP(cov_est='concat', cov_method_params=None, log=None, n_components=4,
  norm_trace=False, rank=None, reg='auto', transform_into='average_power')

	epoch	mean power Fz	mean power POz	label
0	0	1.138202	1.511956	1
1	1	1.194344	1.027722	1
2	2	1.300484	1.081045	1
3	3	1.973375	1.210937	1
4	4	1.269209	1.885737	1
...	...	...	...	...
324	324	1.427216	1.810777	0
325	325	1.239968	1.761608	0
326	326	1.997934	1.459444	0
327	327	1.456173	1.968438	0
328	328	2.140692	2.612114	0